Contribution
A single-case report that a person with ALS, classified by the authors as in the completely locked-in state (CLIS), produced self-chosen words and sentences through an intracortical BCI. The enabling step was not decoding of an intended movement, which failed, but operant training: the patient heard a tone whose pitch followed the firing rate of one to four channels and learned to raise or lower it, and a held high or low tone answered “yes” or “no” to letters read out by a synthetic voice.
Key insight
Communication came back through a newly built contingency. Nothing pre-existing in the motor cortex carried the patient’s intent: “when we instructed the patient to attempt or imagine hand, tongue, or foot movements, we could not detect consistent responses”. What worked was an arbitrary neural variable made audible and reinforced, with an “additional reward sound” when he held the tone on target. The speller itself was the one he had used while he could still move his eyes, so what he carried into CLIS was a learned task, re-attached to a new output.
Assumptions
- CLIS by the absence of communication. “no universally accepted clinical definition exists to distinguish LIS from CLIS; the current standard criteria to differentiate LIS from CLIS is the presence or absence of means of communication” (p. 6). Here, extensive EOG recordings found no usable neuromuscular output, and caretakers and family “denied the existence of any possible reliable communication from February 2019 onwards” (p. 6).
- CLIS was recent. “we conclude based on our reported measurements that the patient described was in a CLIS a few weeks before and also after implantation” (p. 6). The last successful EOG-BCI communication was in February 2019; implantation was in March 2019 (p. 7).
- Not excluded. “even more sensitive measurements of somatic-motor control could reveal some form of volitional control, which would render the diagnostic statement of CLIS at least for this case inaccurate” (p. 6). Post-hoc EOG analysis still found significant yes/no differences in eye-movement features after he could no longer use the EOG system (p. 6).
- Intelligibility rated by authors. Speller output was rated by three of the authors (UC, IV, JZ); an intelligible phrase is one whose meaning “the family or experimenter identified and agreed upon” (p. 8).
Key results
- History before implantation (Methods, p. 7). Diagnosed August 2015 (progressive muscular atrophy, a non-bulbar ALS variant); lost speech and walking by end of 2015; ventilated since July 2016. Used an eye-tracker from August 2016 until August 2017, then a family paper-based scheme on eye movements. He “anticipated complete loss of eye control and asked for an alternative communication system”. From February 2018 he used a non-invasive EOG-based BCI with auditory letter selection (Tonin et al. 2020), “until February 2019”. He gave informed consent to surgery “using his eye movements for confirmation”, before he “transitioned into CLIS” (p. 7).
- Implantation and failure of direct decoding (p. 2). Two 8×8 arrays in hand-area primary motor cortex and supplementary motor area, March 2019. Eye-movement strategy, attempted and imagined movement: no classifiable signal. Passive movement of his fingers, thumb and wrist evoked consistent modulation.
- Neurofeedback (pp. 2–4). From day 86, auditory feedback (120–480 Hz tone mapped from normalized spike rate). He modulated the tone “on his first attempt on day 86”, matched targets “for the first time on day 98”, and spelled from day 106. In the 281 neurofeedback blocks preceding speller blocks (days 123–462), 4936 of 5700 trials (86.6%) were correct; median pre-speller accuracy 90.0%, minimum 50.0% (chance); below 80% in 17.1% of sessions. Firing rates varied considerably within a day, “necessitating manual recalibration throughout the day”.
- Spelling (pp. 4–5). Sessions on 135 days; spelling attempted on 107, intelligible output on 44 (Fig. 4); not attempted on 28 for failing the neurofeedback criterion. Intelligible messages: 5747 characters in 5338 minutes, 1.08 characters per minute (range 0.2–5.1); information transfer 5.2 bits/minute on average. Spearman correlation of neurofeedback accuracy with speller intelligibility 0.282. First copy spelling: his own, his son’s and his wife’s names.
- What he wrote (pp. 4–6). Care requests (“head always straight”, “everybody must use gel on my eye more often”), food, music (“I would like to listen to the album by Tool loud”), suggestions for the speller, and messages to his four-year-old son (“Do you want to watch Disney’s Robin Hood with me”).
- Vision (pp. 2, 7). “With ALS progression, the patient lost the ability to open his eyes voluntarily as well as visual acuity”; “the patient reported low visual acuity caused by the drying of the cornea”. The system is “independent of vision”. Hearing was intact enough for the auditory speller; no audiometry is reported.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | A patient without usable eye or muscle output spelled self-generated sentences through an intracortical BCI | moderate | one patient; 44/107 days intelligible; ratings by authors; data and code deposited |
| C2 | The patient was in CLIS | moderate | EOG and family report; authors concede more sensitive measures might find residual control; post-hoc EOG differences remained |
| C3 | Neurofeedback, not movement decoding, made control possible | moderate | attempted/imagined movement failed for 86 days; feedback succeeded on first attempt |
| C4 | Brain-based volitional communication is possible in CLIS, upending extinction and cognitive-dysfunction hypotheses | weak to moderate | inference from C1–C2 for one recently transitioned patient with prior BCI training |
| C5 | Lack of somatic feedback, disease or arousal changes may explain weak, intermittent control | weak | discussion, no test |
Concepts
- Spike rate metric (SRM). Threshold crossings counted in 50 ms bins, averaged over the past second, per channel; normalized, mixed over selected channels and mapped to tone frequency.
- Feedback with / without reward. Tone-matching blocks; in the second, holding the tone on target for a set time triggers a 250 ms reward sound.
- CLIS. As above: by the authors’ account, defined clinically by the absence of any means of communication.
Connections
It is the test case for Kübler and Birbaumer’s extinction hypothesis (LIT-1199, read in NOTE-948), whose senior author is senior author here. It continues the group’s EOG-based communication work with the same patient (Tonin et al. 2020, cited as ref. 22; the Methods’ “Tonin et al.21” appears to be a misnumbering, since ref. 21 is Kübler and Birbaumer) and contrasts itself with intracortical BCIs in people who kept some muscle control (Jarosiewicz, Pandarinath, Willett, Vansteensel). The decision not to change the decoder rests on learning principles (Birbaumer, Ruiz and Sitaram 2013).
Bearing on the record
- The extinction hypothesis, in the authors’ words. Introduction
(p. 2): hypotheses for CLIS failure range “from extinction of
intentions21 related to protracted loss of sensory input and motor
output, cognitive dysfunction, particularly when it occurs in
association with fronto-temporal degeneration. A successful
demonstration of any BCI enabling an individual without reliable
eye-movement control and with eyes closed (CLIS) to form a complete
sentence would upend these hypotheses”. Discussion (p. 6): “The present
BCI communication demonstrates that an individual unable to move for
protracted periods is capable of meaningful communication.” They treat
the case as refuting it. Three things weigh against that:
- The paraphrase drifts. The 2008 hypothesis was loss of perceived response–outcome contingency; here it is “extinction of intentions … related to protracted loss of sensory input and motor output”, which adds sensory input the 2008 paper said was intact.
- The case fits the hypothesis’s own proviso. Kübler and Birbaumer predicted that “learning BCI-control or any other contingency before onset of CLIS, which can be transferred from LIS to CLIS should prevent extinction”. This patient used an EOG BCI with an auditory speller for a year before CLIS, and “The speller system described here … is identical to the one used previously when he was still in control of eye movements” (p. 8). So he is not the 2008 refuter, a patient who learns while already in CLIS, but the 2008 prediction confirmed.
- CLIS was weeks old, not “protracted”. By the authors’ own dating he was in CLIS “a few weeks before” implantation, and training began on day 86. That is about four months without communication, far from the “longer time periods in CLIS” the hypothesis concerns. The authors themselves keep the mechanism in play: they did not change the decoder because “such a substantial change in the procedure might impede or extinguish the successful control” (p. 6), and they suggest “Lack of any somatic sensory feedback, especially that from muscles, might impede voluntary modulation of neural activity” (p. 7).
- CASE-107. “Chaudhary et al.’s patient communicated only once neurofeedback restored a contingency between intention and outcome” is confirmed by the reading, with two additions: he had a prior BCI history in LIS (above), and the restored contingency was not with an intended movement but with an arbitrary neural variable, after 86 days in which instructed attempts and imagery produced nothing classifiable. The CASE’s “Extinguished or dormant” bullet says the patient “fits the second reading”; on the reading, he fits a third, which the 2008 paper itself allows: a contingency learned before CLIS and kept in use. On the cautions: the article as published carries no editor’s note, correction, expression of concern or retraction (nature.com and Crossref, 2026-10-16). The retractions and outside cautions the LIT reports are not mentioned in it and were not checked here.
- QUESTION-121. It states that hearing survives in CLIS “since Chaudhary et al.’s patient was trained on auditory feedback”. Right for this patient, and vision was failing (corneal drying, eyes closed). The question’s two things do come apart here: the patient kept receiving unselected evidence (people talking to him, his son, music he asked for) through the period without communication, while the contingency was what had to be rebuilt.
- CLAIM-342. The patient’s training loop is anchored the way the claim wants: target tones are set by the experimenter, so his attempt is scored against something he did not select. The observers’ scoring of his spelling is weaker: intelligibility was rated by three authors, and a phrase counted as intelligible when “the family or experimenter identified and agreed upon their meaning” (p. 8). The question paradigm with known answers is mentioned in the Methods but its results are not reported in the main text.
- CLAIM-339 and CLAIM-337. The channel that carried his judgements is an operant contingency on motor-cortex firing, reinforced by a reward tone. That is action in active inference’s sense at its most inward: no movement, no imagined movement, but an output the patient drives and hears back. It supports the CASE’s asymmetry: evidence of lucidity here could only arrive through a contingency the patient controls.
- CLAIM-341. The content of what he wrote, a judgement about what he wants for his son and for supper, is lucid judgement after months without a working channel. It is not judgement with no action at all, since producing it required the rebuilt contingency. Defeater (b) is not tested.
- For the anthology. No instruction for machine-learning practice. An anthology topic on brain–computer interfaces could hold it; the record holds it for the extinction question.
Limitations
- One patient, in CLIS for weeks rather than years, with prior BCI training; residual control below the EOG’s sensitivity not excluded.
- Control was intermittent: intelligible output on 44 of 107 attempted days, with daily manual recalibration.
- Intelligibility judged by authors and family; no blinded scoring of speller output is reported.
- Supplementary materials, including the EOG analysis behind the CLIS classification, were not read for this note.
Open questions
- Would a patient in CLIS for years, with no prior BCI, learn the same neurofeedback task? That is the 2008 hypothesis’s real test.
- Did the patient’s control depend on the reward tone (operant) or would plain feedback suffice?